Measuring What Matters

Marianela Lopez

CEO & Founder Empujón Educativo

Marianela López is the founder and CEO of Empujón Educativo, an initiative born out of her experience supporting her son, Amador, and aimed at transforming education through a deep understanding of how each student learns. She leads the development of solutions that integrate neuroscience, behavioral sciences, and ethical, responsible artificial intelligence to turn data into pedagogical insight. Her work focuses on language, learning, and inclusion, guided by a core conviction: technology can identify patterns, but only a human perspective can construct educational meaning.

AI & EDUCATION — CHAPTER 2

Measuring What Matters

The number that lies

A bilingual private school in Buenos Aires scores three times higher than a local public school on standardized reading-fluency tests. Drill beneath the aggregate score, and the picture shifts. The bilingual school’s students read faster, but they make more errors. The public-school students are slower and more precise. Whether that gap reflects genuine differences in learning quality, or differences in how students have been prepared for this specific type of assessment, is a question that the score alone cannot answer. What it suggests, at minimum, is that the relationship between what gets measured and what is actually being learned is more complicated than the available data implies.

This kind of complexity is not unique to reading tests. It runs, in different forms, through most of the data that Latin American schools currently generate—and it has particular relevance now, as artificial intelligence enters educational settings. If the data encodes the wrong questions, the systems will amplify wrong answers—consistently, at scale, and without anyone noticing.

The region’s education challenge is not only about connectivity, resources, teacher training—though it is all of those things too. It is, at its foundation, a measurement challenge. And measurement is never a neutral act. Choosing which metrics to collect is choosing which version of human potential a society decides to value.

Language as a lever of exclusion

1. Bilingual education and inclusion

Data is collected on language, and language in Latin American education is not only a medium of instruction. It is a mechanism of selection. Who speaks which variety of which language, with what accent, in what register, determines—often before any content is taught—how a student is positioned within the educational system and how their performance will be read.

The most visible version of this dynamic plays out in bilingual schools. In Argentina and across several other countries in the region, private bilingual institutions have become the dominant model for upper-income urban families, and English instruction is increasingly being integrated into public-school curricula as well. For most students, this expansion represents opportunity. For students with learning differences, it frequently operates as an exclusion mechanism. 

Choosing which metrics to collect is choosing which version of human potential a society decides to value.

When these students struggle in a bilingual setting, the standard clinical recommendation is to simplify their educational path by removing the second language. In practice, this means leaving the school or the track, which means leaving the peer group, the social environment, and in many cases, the educational trajectory that was available to them. The second language was supposed to open doors. For these students, it closes one.

2. Dialectal equity as a design benchmark

The bilingual dynamic is, in some ways, the most legible version of a problem that runs much deeper and much wider across the region. Spanish itself is not a single, homogeneous system. It encompasses substantial variation in phonology, vocabulary, rhythm, and syntax across countries, regions, and communities. 

The problem is that most AI tools used in educational settings, such as speech recognition models, reading-fluency assessment platforms, and language-processing tools are typically trained on standardized or Castilian Spanish. When these models encounter dialectal variation, they code it as deviation. The student is flagged as misreading. The assessment records an error that is not there. At the level of a single session, this is an inaccuracy. At the level of a regional education system using automated reading assessments to make decisions about learning support, grade progression, and resource allocation, it is a structural bias.

This is what practitioners in the field mean when they speak of dialectal equity as a design objective rather than an afterthought. Building AI tools that work equitably for Latin American learners requires training on the actual diversity of spoken Spanish across the region—including the varieties that are underrepresented in existing datasets precisely because the communities that speak them have been underrepresented in everything else. A model that penalizes regional speech as error does not just produce inaccurate data. It encodes, in the architecture of the assessment itself, the same hierarchy that the educational system is supposed to be working against.

That hierarchy does not stop at language. It structures access to the system at every level.

Who gets filtered out—and when

In recent years, Latin America has made genuine progress on educational access at the primary level. Enrollment rates are near-universal in most countries, and the political commitment to keeping children in school through primary school has produced measurable results. The problem is what happens next.

Secondary education remains the region’s most structurally exclusive level. According to the Inter-American Development Bank, roughly one in three young people in Latin America does not reach upper secondary school, and of those who do enroll, close to half do not graduate. The transition between levels is where inequality concentrates, and where three groups bear the heaviest burden:

• Indigenous students, who enroll in higher education at rates 32% below their non-indigenous peers; 
• students with disabilities, who attend secondary school at rates roughly 10% lower than those without; 
• and rural students, particularly young men, whose entry into the labor market is structured by economic necessity rather than choice.

But there is a fourth, less visible group whose exclusion begins earlier and is harder to track: neurodiverse students. Approximately 15 to 20 percent of the global population is neurodivergent—a proportion that, applied to Latin America’s population of around 660 million, represents well over 100 million people. Many of them are currently filtered out of the formal education system before their capabilities are ever meaningfully assessed.

Building AI tools that work equitably for Latin American learners requires training on the actual diversity of spoken Spanish across the region—including the varieties that are underrepresented in existing datasets precisely because the communities that speak them have been underrepresented in everything else.

In fact, in several Latin American countries, students with specific learning needs who complete primary school under adapted curricula do so without credentials that secondary institutions recognise. The policy commitment to universal primary enrollment was built without redesigning the certification architecture that governs access to the next level.

The result is a structural discontinuity: Students who have been formally included in one part of the system find themselves formally excluded from the next.

The personalization paradox

AI does not resolve this gap, but it changes what is observable within it. The volume and granularity of data it can process allows a kind of attention to individual learning trajectories that a single teacher managing thirty students cannot sustain alone—and that opens, at least in principle, the possibility of genuinely personalized learning.

Personalization in Latin American education is, for now, more hypothesis than practice. The awareness that students do not all learn in the same way has grown significantly among school leaders and teachers. But awareness and implementation are different things. Most classrooms still operate with a single teacher managing thirty or more students, a uniform set of materials, and an assessment system built around the assumption that the same test, administered at the same moment, measures the same thing for everyone.

What makes the transition toward personalization genuinely difficult is not only a question of resources or teacher training, though both are real constraints. It is also a cultural one. Latin American schools are, in most cases, organized around comparative ranking and meritocratic logic. When a student receives differentiated materials or adapted assessments within this environment, the difference is visible to peers, and peers interpret it. In some classrooms, the student receiving support is read as receiving an advantage. In others, as being incapable. Both interpretations create social dynamics that work against the very inclusion the adaptation was designed to support.

This tension surfaced early in field research in the region: In some cases, the first to resist adapted materials were the students themselves, aware of how difference gets coded in a system that rewards uniformity. Changing that dynamic is not a technical problem. It requires shifting the cultural assumptions embedded in how schools understand achievement, how they display it, and what they signal, implicitly or explicitly, about the students who do not conform to a single standard. That shift is happening, in some schools more than others, but it is slow, uneven, and largely invisible to the policy frameworks that are supposed to support it.

The data gap

That invisibility has a structural explanation. The data that Latin American schools generate is, in most cases, too thin to surface what is actually happening in classrooms—let alone to support the kind of AI-assisted personalization that is regularly promised in education technology discussions.

Grades and attendance—the two data points that most school management systems collect systematically—capture almost nothing about how individual students learn, where specific difficulties arise, what cognitive or emotional factors shape performance on a given day, or how a student’s profile changes over time in response to different teaching approaches. Training AI on this data produces a system that cannot meaningfully personalize learning, identify a student’s strengths alongside their difficulties, or support the kind of decision-making that inclusive education actually requires.

The interoperability problem compounds this. When a student changes schools—which is common among the region’s most vulnerable populations—their learning history typically does not follow them. Educational platform data belongs to the platform, not to the student or the family, and platforms do not communicate with each other. A student who spent three years developing a documented learning profile in one school arrives at the next as an unknown quantity, and the process of understanding them begins again from scratch.

Changing that dynamic is not a technical problem. It requires shifting the cultural assumptions embedded in how schools understand achievement, how they display it, and what they signal, implicitly or explicitly, about the students who do not conform to a single standard.

The regulatory environment offers limited protection. There are no unified data governance frameworks across the region, and the conversations about AI in education are still largely happening at the design level—in policy documents and technical working groups—rather than in schools. Initiatives like Argentina’s Paideia, which evaluates and catalogs AI tools for educational use at the national level, are a step toward institutional scaffolding. But between a tool being validated at the policy level and that tool having a meaningful impact in a classroom is a distance that most current systems have not yet learned to bridge.

Five years out

The trajectory for AI in Latin American education over the next five years will be shaped less by the capabilities of the technology than by whether the region resolves a set of upstream problems that the technology cannot resolve on its own.

Connectivity remains the most basic. In Argentina and Mexico, schools serving thousands of students operate with a handful of computers or none at all. In some rural settings, students share family phones to access learning platforms. Without basic infrastructure, conversations about AI-assisted personalization are conversations that exclude most of the students who would benefit most from it.

Uruguay’s Ceibal program, launched in 2007 as the world’s first national one-to-one laptop initiative and now encompassing universal Wi-Fi coverage in public schools, demonstrates what sustained public investment in digital infrastructure can produce. It also demonstrates how long it takes: nearly two decades of continuous implementation to arrive at a position where more sophisticated applications of technology can be meaningfully deployed.

Beyond connectivity, what will determine the educational trajectory of the region is whether the measurement problem gets taken seriously. If AI tools are adopted at scale while the underlying metrics remain unreformed then AI will accelerate the existing system rather than improve it. The alternative requires treating data quality, data governance, and assessment design as public infrastructure questions, not as technical details to be resolved by individual vendors. 

Artificial intelligence is nothing more than a mirror of our existing problems. However, this is not a reason to reject AI in education. It is a reason to think carefully about what AI is being asked to do, and with what data. Technology alone does not create inequalities. It finds them, processes them, and returns them, faster and at a greater scale.

The Work of Empujón Educativo

From a personal problem to a research question

Empujón Educativo began as a workaround. Its founder, Marianela López, a social psychologist and behavioral science researcher, built a simple tool to create accessible versions of her son’s school materials—her son has Asperger syndrome and severe dyslexia, and the standard curriculum was not designed for how he learns. When the school asked to use the tool for other students, what had been a private adaptation became a pedagogical experiment. That was 2023. By early 2024, it had become a research project.

The pivot was driven by two obstacles that quickly became the initiative’s defining concerns. The first was the poverty of available educational data: grades, attendance, and a single reading fluency score per student—enough to train a model on the surface of learning, not on its substance. The second was the absence of any commercial transcription tool capable of accurately decoding the specific errors students make when reading aloud in regionally varied Spanish. Both problems, it turned out, were not incidental to the region’s education challenge. They were central to it.

What the technology actually measures

Empujón’s approach centers on a set of micro-applications designed to capture data that standard school records do not produce. The reading fluency assessment is the clearest example: Rather than recording words per minute, the system transcribes oral reading at the syllabic level, classifying specific error types to build a functional profile of how a student is processing written language. Eye movement and facial micro-expression data are collected alongside reading performance, moving toward a multi-variable picture of the learning process that can inform teaching decisions without generating diagnostic labels.

The transcription model required to make this work does not exist commercially. Empujón Educativo is developing it—trained on regional dialectal variation, designed to distinguish between a reading error and a phonological feature of a student’s local Spanish, and calibrated to avoid encoding linguistic standardization as educational correctness. The team describes this as a model of dialectal equity: The premise that accurate assessment requires recognizing the validity of regional speech rather than measuring students against a norm that, in practice, is not neutral at all.

The initiative also addresses the bilingual gap directly, developing simplified text adaptations, concept maps, and multimedia support for students with learning differences in bilingual settings—tools designed to keep students in mainstream educational environments rather than accepting exclusion as the path of least resistance.

On data as a public good

One of Empujón’s more structurally significant positions is its approach to data ownership. The initiative maintains an open-source, interoperable educational data archive built on the principle that learning histories belong to students and families, not to the platforms that generate them. In a regional context where a student who changes schools typically loses all accumulated data, and where teachers must re-enter the same information into multiple non-communicating systems, this is not a technical choice. It is a political one.

Current scale

Empujón operates across Argentina, Mexico, Peru, and Uruguay, with a deliberate focus on heterogeneous settings—rural and urban, public and private—to ensure its datasets reflect the actual diversity of the region’s learners. Ten school pilots have been completed. The current roadmap targets 50 schools and approximately 50,000 students within a single Argentine province—a sample large enough to begin building regionally grounded reading fluency benchmarks that include, rather than exclude, the students that existing standardized measures were not designed to see.